The Annals of Probability 2018, Vol. 46, No. 6, 3015–3089 https://doi.org/10.1214/17-AOP1226 © Institute of Mathematical Statistics, 2018
PFAFFIAN SCHUR PROCESSES AND LAST PASSAGE PERCOLATION IN A HALF-QUADRANT1
∗ BY JINHO BAIK ,2,GUILLAUME BARRAQUAND†,3,IVA N CORWIN†,4 AND TOUFIC SUIDAN University of Michigan∗ and Columbia University†
We study last passage percolation in a half-quadrant, which we analyze within the framework of Pfaffian Schur processes. For the model with expo- nential weights, we prove that the fluctuations of the last passage time to a point on the diagonal are either GSE Tracy–Widom distributed, GOE Tracy– Widom distributed or Gaussian, depending on the size of weights along the diagonal. Away from the diagonal, the fluctuations of passage times fol- low the GUE Tracy–Widom distribution. We also obtain a two-dimensional crossover between the GUE, GOE and GSE distribution by studying the mul- tipoint distribution of last passage times close to the diagonal when the size of the diagonal weights is simultaneously scaled close to the critical point. We expect that this crossover arises universally in KPZ growth models in half-space. Along the way, we introduce a method to deal with diverging cor- relation kernels of point processes where points collide in the scaling limit.
CONTENTS 1. Introduction ...... 3016 1.1. Previous results in (half-space) LPP ...... 3017 1.2. Previous methods ...... 3018 1.3. Our methods and new results ...... 3018 1.4.OthermodelsrelatedtoKPZgrowthinahalf-space...... 3020 1.5.Mainresults...... 3020 Outline of the paper ...... 3024 2. Fredholm Pfaffian formulas ...... 3024 2.1.GUETracy–Widomdistribution...... 3024 2.2. Fredholm Pfaffian ...... 3025 2.3.GOETracy–Widomdistribution...... 3026 2.4.GSETracy–Widomdistribution...... 3027
Received July 2016; revised August 2017. 1Supported by NSF PHY11-25915. 2Supported in part by NSF Grant DMS-1361782. 3Supported in part by the LPMA UMR CNRS 7599, Université Paris-Diderot–Paris 7 and the Packard Foundation through I. Corwin’s Packard Fellowship. 4Supported in part by the NSF through DMS-1208998, a Clay Research Fellowship, the Poincaré Chair and a Packard Fellowship for Science and Engineering. MSC2010 subject classifications. Primary 60K35, 82C23; secondary 60G55, 05E05, 60B20. Key words and phrases. Last passage percolation, KPZ universality class, Tracy–Widom distri- butions, Schur process, Fredholm Pfaffian, phase transition. 3015 3016 BAIK, BARRAQUAND, CORWIN AND SUIDAN
2.5.Crossoverkernels...... 3029 3.PfaffianSchurprocess...... 3031 3.1. Definition of the Pfaffian Schur process ...... 3031 3.1.1. Partitions and Schur positive specializations ...... 3031 3.1.2. Useful identities ...... 3033 3.1.3. Definition of the Pfaffian Schur process ...... 3033 3.2.PfaffianSchurprocessesindexedbyzig-zagpaths...... 3036 3.3.DynamicsonPfaffianSchurprocesses...... 3037 3.4.Thefirstcoordinatemarginal:Lastpassagepercolation...... 3041 4. Fredholm Pfaffian formulas for k-pointdistributions...... 3044 4.1.Pfaffianpointprocesses...... 3045 4.2.Half-spacegeometricweightLPP...... 3046 4.3.Deformationofcontours...... 3048 4.4. Half-space exponential weight LPP ...... 3050 5.ConvergencetotheGSETracy–Widomdistribution...... 3058 5.1.AformalrepresentationoftheGSEdistribution...... 3059 5.2. Asymptotic analysis of the kernel Kexp ...... 3068 5.3.Conclusion...... 3073 6.Asymptoticanalysisinotherregimes:GUE,GOE,crossovers...... 3074 6.1. Phase transition ...... 3074 6.2.GOEasymptotics...... 3076 6.3.GUEasymptotics...... 3078 6.4.Gaussianasymptotics...... 3084 6.5. Convergence of the symplectic-unitary transition to the GSE distribution ...... 3086 Acknowledgments...... 3087 References...... 3087
1. Introduction. In this paper, we study last passage percolation in a half- quadrant of Z2. We extend all known results for the case of geometric weights (see discussion of previous results below) to the case of exponential weights. We study in a unified framework the distribution of passage times on and off diagonal, and for arbitrary boundary condition. We do so by realizing the joint distribution of last passage percolation times as a marginal of Pfaffian Schur processes. This allows us to use powerful methods of Pfaffian point processes to prove limit theorems. Along the way, we discuss an issue arising when a simple point process converges to a point process where every point has multiplicity two, which we expect to have independent interest. These results also have consequences about interacting particle systems—in particular the TASEP on positive integers and the facilitated TASEP—that we discuss in [2].
DEFINITION 1.1 (Half-space exponential weight LPP). Let (wn,m)n≥m≥0 be a sequence of independent exponential random variables5 with rate 1 when n ≥ m +
5The exponential distribution with rate α ∈ (0, +∞), denoted E(α), is the probability distribution −αx on R>0 such that if X ∼ E(α), P(X > x) = e . PFAFFIAN SCHUR PROCESSES AND LPP IN A HALF-QUADRANT 3017
FIG.1. LPP on the half-quadrant. One admissible path from (1, 1) to (n, m) is shown in dark gray. H(n,m) is the maximum over such paths of the sum of the weights wij along the path.
1 and with rate α when n = m. We define the exponential last passage percolation (LPP) time on the half-quadrant, denoted H(n,m), by the recurrence for n ≥ m, max H(n− 1,m),H(n,m− 1) if n ≥ m + 1, H(n,m)= wn,m + H(n,m− 1) if n = m with the boundary condition H(n,0) = 0.
It is useful to notice that the model is equivalent to a model of last passage percolation on the full quadrant where the weights are symmetric with respect to the diagonal (wij = wji).
REMARK 1.2. If one imagines that the light gray area in Figure 1 corresponds to the percolation cluster at some time, its border (shown in black in Figure 1) evolves as the height function in the Totally Asymmetric Simple Exclusion process on the positive integers with open boundary condition, so that all our results could be equivalently phrased in terms of the latter model.
1.1. Previous results in (half-space) LPP. The history of fluctuation results in last passage percolation goes back to the solution to Ulam’s problem about the asymptotic distribution of the size of the longest increasing subsequence in a uni- formly random permutation [4]. A proof of the nature of fluctuations in exponential distribution LPP is provided in [29]. On a half-space, [5–7] studies the longest in- creasing subsequence in random involutions. This is equivalent to a half-space LPP problem since for an involution σ ∈ Sn, the graph of i → σ(i) is symmetric with respect to the first diagonal. Actually, [5–7] treat both the longest increasing sub- sequences of a random involution and symmetrized LPP with geometric weights. That work contains the geometric weight half-space LPP analogue of Theorem 1.3. 3018 BAIK, BARRAQUAND, CORWIN AND SUIDAN
The methods used therein only worked when restricted to the one point distribu- tion of passage times exactly along the diagonal. Further work on half-space LPP was undertaken in [38], where the results are stated there in terms of the discrete polynuclear growth (PNG) model rather than the equivalent half-space LPP with geometric weights. The framework of [38] allows to study the correlation kernel corresponding to multiple points in the space direction, with or without nucleations at the origin. In the large scale limit, the authors performed a nonrigorous critical point derivation of the limiting correlation kernel in certain regimes and found Airy-type process and various limiting crossover kernels near the origin, much in the same spirit as our results (see the discussion about these results in Section 1.3).
1.2. Previous methods. A key property in the solution of Ulam’s problem in [4] was the relation to the RSK algorithm which reveals a beautiful algebraic struc- ture that can be generalized using the formalism of Schur measures [34] and Schur processes [35]. RSK is just one of a variety of Markov dynamics on sequences of partitions which preserves the Schur process [13, 17]. Studying these dynamics gives rise to a number of interesting probabilistic models related to Schur pro- cesses. In the early works of [4, 6], the convergence to the Tracy–Widom distribu- tions was proved by Riemann–Hilbert problem asymptotic analysis. Subsequently, Fredholm determinants became the preferred vehicle for asymptotic analysis. In a half-space, [5] applied RSK to symmetric matrices. Subsequently, [37](see also [26]) showed that geometric weight half-space LPP is a Pfaffian point process (see Section 4.1) and computed the correlation kernel. Later, [21] defined Pfaffian Schur processes generalizing the probability measures defined in [37]and[38]. The Pfaffian Schur process is an analogue of the Schur process for symmetrized problems.
1.3. Our methods and new results. We consider Markov dynamics similar to those of [17] and show that they preserve Pfaffian Schur processes. This enables us to show how half-space LPP with geometric weights is a marginal of the Pfaffian Schur process. We then apply a general formula giving the correlation kernel of Pfaffian Schur processes [21] to study the model’s multipoint distributions. We degenerate our model and formulas to the case of exponential weight half-space LPP (previous works [5–7, 38] only considered geometric weights). We perform an asymptotic analysis of the Fredholm Pfaffians characterizing the distribution of passage times in this model, which leads to Theorems 1.3 and 1.4.Wealso study the k-point distribution of passage times at a distance O(n2/3) away from the diagonal, and introduce a new two-parameter family of crossover distributions when the rate of the weights on the diagonal are simultaneously scaled close to their critical value. This generalizes the crossover distributions found in [25, 38]. The asymptotics of fluctuations away from the diagonal were already studied in [38] for the PNG model on a half-space (equivalently the geometric weight half- space LPP). The asymptotic analysis there was nonrigorous, at the level of studying PFAFFIAN SCHUR PROCESSES AND LPP IN A HALF-QUADRANT 3019
the behavior of integrals around their critical points without controlling the decay of tails of integrals. The proof of the first part of Theorem 1.3 (in Section 5) required a new idea which we believe is the most novel technical contribution of this paper: When α>1/2, H(n, n) is the maximum of a simple Pfaffian point process which con- verges as n →∞to a point process where all points have multiplicity 2. The limit of the correlation kernel does not exist6 as a function (it does have a formal limit involving the derivative of the Dirac delta function). Nevertheless, a careful re- ordering and nontrivial cancellation of terms in the Fredholm Pfaffian expansions allows us to study the law of the maximum of this limiting point process, and ulti- mately find that it corresponds to the GSE Tracy–Widom distribution. We actually provide a general scheme for how this works for random matrix type kernels in which simple Pfaffian point processes limit to doubled ones (see Section 5.1). Theorem 5.3 of [38] derives a formula for certain crossover distributions in half- space geometric LPP. This crossover distribution, introduced in [25], depends on a parameter τ (which is denoted η in our Theorem 1.8) and it corresponds to a natu- ral transition ensemble between GSE when τ → 0 and GUE when τ →+∞.Al- though the collision of eigenvalues should occur in the GSE limit of this crossover, this point was left unaddressed in previous literature and the convergence of the crossover kernel to the GSE kernel was not proved in [25] (see our Proposi- tion 6.12). The formulas for limiting kernels, given by [25], (4.41), (4.44), (4.47) or [38], (5.37)–(5.40), make sense so long as τ>0. Notice a difference of integration domain between [25], (4.44), and [38], (5.39). Our Theorem 1.8 is the analogue of [38], Theorem 5.3, for half-space exponential LPP and we recover the exact same kernel. When τ = 0, the formula for I4 in [ 38], (5.39), involves divergent inte- 7 grals, though if one uses the formal identity R Ai(x + r)Ai(y + r)dr = δ(x − y) then it is possible to rewrite the kernel in terms of an expression involving the derivative of the Dirac delta function, and the expression formally matches with the kernel K∞ introduced in Section 5.1. From that (formal) kernel, it is nontrivial to match the Fredholm Pfaffian with a known formula for the GSE distribution, this matching does not seem to have been made previously in the literature and we explain it in Section 5.1. In random matrix theory, a similar phenomenon of collisions of eigenvalues oc- curs in the limit from the discrete symplectic ensemble to the GSE [22]. However, [22] used averages of characteristic polynomials to characterize the point process, and computed the correlation kernels from the characteristic polynomials only af- ter the limit, so that collisions of eigenvalues were not an issue. A random matrix ensemble which crosses over between GSE and GUE was studied in [25].
6For a nonsimple point process, the correlation functions—Radon–Nikodym derivatives of the factorial moment measures—generically do not exist (see Definitions in Section 4.1). 7The saddle-point analysis in the proof of Theorem 5.3 in [38] is valid only when τ>0. Indeed, [38], (5.44), requires τ1 + τ2 >η1 + η2 and one needs that η1,η2 > 0 as in the proof of [38], Theorem 4.2. 3020 BAIK, BARRAQUAND, CORWIN AND SUIDAN
1.4. Other models related to KPZ growth in a half-space. A positive temper- ature analogue of LPP—directed random polymers—has also been studied. The log-gamma polymer in a half-space (which converges to half-space LPP with ex- ponential weights in the zero temperature limit) is considered in [33]wherean exact formula is derived for the distribution of the partition function in terms of Whittaker functions. Some of these formulas are not yet proved and presently there has not been a successful asymptotic analysis preformed of them. Within physics, [28]and[14] studies the continuum directed random polymer (equivalently the KPZ equation) in a half-space with a repulsive and refecting (respectively) barrier and derives GSE Tracy–Widom limiting statistics. Both works are mathematically nonrigorous due to the ill-posed moment problem and certain unproved forms of the conjectured Bethe ansatz completeness of the half-space delta Bose gas. On the side of particle systems, half-space ASEP has first been studied in [41], but the resulting formulas were not amenable to asymptotic analysis. More recently, [9] derives exact Pfaffian formulas for the half-space stochastic six-vertex model, and proves GOE asymptotics for the current at the origin in half-space ASEP, for a cer- tain boundary condition. The exact formulas involve the correlation kernel of the Pfaffian Schur process and the asymptotic analysis in [9]usessomeoftheideas developed here.
1.5. Main results. We now state the limit theorems that constitute the main results of this paper. They involve the GOE, GSE and GUE Tracy–Widom distri- butions, respectively, characterized by the distribution functions FGOE, FGSE and FGUE given in Section 2 and the standard Gaussian distribution function denoted by G(x).
THEOREM 1.3. The last passage time on the diagonal H(n, n) satisfies the following limit theorems, depending on the rate α of the weights on the diagonal: 1. For α>1/2, H(n, n) − 4n lim P